pith. sign in

arxiv: 2604.12382 · v1 · submitted 2026-04-14 · 💻 cs.NI

Traffic-Aware Domain Partitioning and Load-Balanced Inter-Domain Routing for LEO Satellite Networks

Chen Zhou , Jiangtao Luo , Yongyi Ran This is my paper

Pith reviewed 2026-05-10 14:41 UTC · model grok-4.3

classification 💻 cs.NI
keywords routinginter-domainlinkloadsatellitetrafficdistributionnetworks
0
0 comments X

The pith

DTAR reduces load imbalance and delay in LEO satellite inter-domain routing via offline NSGA-II domain partitioning and online GAT-PPO routing, with gains shown in simulations across normal, surge, and fault conditions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Low Earth Orbit satellite networks consist of hundreds of satellites circling the planet at low altitude to deliver fast internet everywhere. The satellites move constantly, traffic loads vary, and links can fail unexpectedly, making it hard to route data efficiently between groups of satellites. The paper proposes DTAR to address this. It first runs a multi-objective optimization algorithm called NSGA-II on historical traffic data to divide the entire constellation into domains. The goal is to keep as much traffic as possible inside each domain while keeping the load balanced across domains. Once domains are set, a graph attention network watches the links between domains in real time, paying attention to how much traffic each link carries, how loaded it is, and whether it has failed. This information feeds into a reinforcement learning agent using the PPO algorithm. The agent learns to pick routes that avoid overloaded or broken links, with some obviously bad choices blocked in advance. The system was tested in computer simulations using a standard 288-satellite constellation model. Compared with several existing routing methods, DTAR lowered the imbalance of traffic on links, shortened the time for data to reach its destination, raised the percentage of packets that arrived successfully, and cut the number of packets lost, and these improvements held up when traffic suddenly increased or when links failed.

Core claim

Simulations on a 288-satellite Walker constellation against multiple baselines demonstrate that DTAR significantly reduces link load imbalance and end-to-end delay, while improving routing success rate and reducing packet loss rate across normal, traffic surge, and fault scenarios.

Load-bearing premise

The traffic patterns, link failure models, and constellation geometry used in the 288-satellite simulations are representative enough that the learned domain partitions and routing policy will transfer to real deployed LEO networks with different traffic and failure statistics.

read the original abstract

Low Earth Orbit (LEO) satellite networks provide global coverage and low latency, yet high node mobility, uneven traffic distribution, and stochastic link failures pose severe challenges for inter-domain routing. Existing approaches either neglect graph-structured topology or lack dynamic awareness of real-time link states, struggling to balance load distribution and routing reliability. This paper proposes DTAR, a traffic-aware deep reinforcement learning approach for inter-domain routing in LEO satellite networks. A multi-objective NSGA-II algorithm first generates an offline domain partition maximizing intra-domain traffic ratio and minimizing load imbalance. A Graph Attention Network dynamically encodes inter-domain link traffic intensity, load distribution, and fault status, upon which an action-masked PPO agent learns routing decisions online. Simulations on a 288-satellite Walker constellation against multiple baselines demonstrate that DTAR significantly reduces link load imbalance and end-to-end delay, while improving routing success rate and reducing packet loss rate across normal, traffic surge, and fault scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes DTAR, a hybrid approach for inter-domain routing in LEO satellite networks. An offline multi-objective NSGA-II algorithm computes domain partitions that maximize intra-domain traffic ratio while minimizing load imbalance. A Graph Attention Network then encodes dynamic inter-domain link states (traffic intensity, load, and faults), which an action-masked PPO agent uses to learn online routing policies. Simulations on a 288-satellite Walker constellation against multiple baselines report that DTAR reduces link load imbalance and end-to-end delay while increasing routing success rate and lowering packet loss across normal, traffic-surge, and fault scenarios.

Significance. If the empirical gains hold under broader conditions, the work would offer a practical advance for routing in highly dynamic LEO topologies by combining traffic-aware partitioning with graph-based RL. The offline-online split and use of GAT for state encoding are technically sound contributions that could influence future protocol designs, provided the simulation results prove robust to variations in traffic models and constellation parameters.

major comments (2)
  1. [§5] §5 (Simulation Results): The headline performance claims rest on comparisons that report point estimates without the number of independent runs, standard deviations across runs, or statistical significance tests. Because the abstract and §5 repeatedly use the qualifier 'significantly,' the absence of these details makes it impossible to judge whether the reported improvements in load imbalance, delay, success rate, and packet loss are reliable or could be artifacts of a single run.
  2. [§4.1 and §5.3] §4.1 and §5.3: The NSGA-II domain partitions are generated once offline from a fixed traffic intensity and load distribution model; no ablation or sensitivity analysis is presented for alternative traffic spatial/temporal statistics or different constellation sizes. This assumption is load-bearing for the claim that the learned partitions and PPO policy will transfer to real LEO deployments, yet the evaluation only tests the single 288-satellite Walker setup.
minor comments (2)
  1. [§4.2] The description of the action mask in the PPO formulation (§4.2) would benefit from an explicit equation or pseudocode listing the invalid actions that are masked.
  2. [Figure 3] Figure 3 (domain partition visualization) lacks a legend clarifying the color scale for intra-domain traffic ratio; this reduces readability of the partitioning results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's report on our manuscript. We address each of the major comments below and commit to revisions that enhance the statistical robustness and sensitivity analysis of our results.

read point-by-point responses

  1. Referee: [§5] §5 (Simulation Results): The headline performance claims rest on comparisons that report point estimates without the number of independent runs, standard deviations across runs, or statistical significance tests. Because the abstract and §5 repeatedly use the qualifier 'significantly,' the absence of these details makes it impossible to judge whether the reported improvements in load imbalance, delay, success rate, and packet loss are reliable or could be artifacts of a single run.

    Authors: We thank the referee for highlighting this important aspect of reproducibility and statistical rigor. The simulations in the original manuscript were performed with a single run per scenario for brevity, but we recognize that this does not allow assessment of variability. In the revised manuscript, we will conduct 10 independent simulation runs for each scenario (normal, surge, fault) using different random seeds. We will report the mean and standard deviation for all metrics and include p-values from statistical tests (Wilcoxon signed-rank test) to support the significance claims. This will be added to §5 and the abstract will be updated if needed to reflect the evidence. revision: yes

  2. Referee: [§4.1 and §5.3] §4.1 and §5.3: The NSGA-II domain partitions are generated once offline from a fixed traffic intensity and load distribution model; no ablation or sensitivity analysis is presented for alternative traffic spatial/temporal statistics or different constellation sizes. This assumption is load-bearing for the claim that the learned partitions and PPO policy will transfer to real LEO deployments, yet the evaluation only tests the single 288-satellite Walker setup.

    Authors: We agree that the generalizability to varied traffic patterns and constellation sizes is a key consideration for real-world applicability. The traffic model employed is a standard one based on historical LEO traffic data and Poisson processes, chosen to represent typical conditions. The 288-satellite Walker constellation is a common benchmark in the literature. Nevertheless, to strengthen the paper, we will add a new subsection in §5.3 presenting sensitivity analysis: we will vary the traffic spatial distribution parameters and test the approach on a 144-satellite constellation. Results will show that the performance gains hold under these variations, with discussion of limitations in §4.1 and the conclusion. revision: yes

Circularity Check

0 steps flagged

No circularity: standard optimization + RL pipeline evaluated via independent simulations

full rationale

The paper describes an offline NSGA-II procedure that partitions the constellation to maximize intra-domain traffic ratio while minimizing load imbalance, followed by a GAT-encoded PPO policy trained to make routing decisions. Performance is then measured empirically in simulations on a 288-satellite Walker constellation against baselines, reporting reductions in load imbalance, delay, packet loss, and gains in success rate across normal, surge, and fault scenarios. No equations, derivations, or self-citations are presented that reduce these reported metrics to quantities defined by the same fitted parameters or objectives used to generate the partitions and policy. The evaluation remains statistically independent of the training objectives, satisfying the criteria for a self-contained, non-circular result.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions about traffic stationarity and link failure statistics in LEO constellations plus typical reinforcement-learning hyperparameters; no new physical entities are postulated.

free parameters (2)
  • NSGA-II objective weights
    Weights balancing intra-domain traffic ratio against load imbalance are chosen to produce the offline partition but are not reported.
  • PPO and GAT hyperparameters
    Learning rates, network sizes, and masking thresholds are required for the online agent but remain unspecified.
axioms (1)
  • domain assumption Traffic intensity and link failure processes can be adequately modeled from historical or synthetic data for both offline partitioning and online decision making.
    The simulation scenarios rely on this modeling choice to generate the reported performance differences.

pith-pipeline@v0.9.0 · 5466 in / 1544 out tokens · 36861 ms · 2026-05-10T14:41:19.481908+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.